Our study enrolled patients with HCC after curative resection between January 2010 and December 2015 in our hospital. Basic information was also received from 86 HCC patients who were hospitalized at the Southern Medical University Zhujiang Hospital (external cohort 1) between January 2016 and January 2019. We also enrolled 52 HCC patients from Hunan cancer hospital (external cohort 2) between June 2014 and September 2016. These patient data were used for additional external validation of the proposed model. The inclusion criteria were as follows: (1) Child‒Pugh class A or B; (2) curative liver resection performed; (3) HCC diagnosis based on postoperative pathology; (4) availability of contrast-enhanced MRI scans obtained within 1 month before surgical resection; (5) a maximum diameter of the HCC lesion exceeding 10 mm. The exclusion criteria were as follows: (1) less than 24 months of follow-up; (2) use of other treatments, such as chemotherapy or interventional therapy, before surgery; (3) presence of other malignant tumors; (4) unqualified image artifacts; and (5) incomplete clinical data and follow-up information.

A total of 494 patients (428 men and 66 women; mean age, 51.28 years ± 11.29; range, 18–80 years) were enrolled in our study. For temporally independent testing, patients who underwent surgery prior to April 2015 were assigned to a training cohort (n = 346), and subsequently operated patients were assigned to a testing cohort (n = 148), at a ratio of 7:3. The training cohort was used to construct the models that were then assessed using the testing cohort.

This retrospective study was approved by the Sun Yat-sen University Cancer Center Institutional Review Board (B2021-214-01), and the requirement for obtaining informed consent was waived.

Recurrence-free survival (RFS) and OS served as the study endpoints. All patients would undergo contrast-enhanced computed tomography (CT) or MRI, and test serum alpha-fetoprotein (AFP) levels every 3 months during the first 2 years after hepatectomy, and every 6 months thereafter to determine patient survival status after surgical resection. The data were censored on December 30, 2020. RFS was defined as the time from the date of surgery to the date of first recurrence. OS was defined as the period from the time of HCC surgery to the time of death or last follow-up.

All patients underwent hepatectomy with tumor-margin-negative resection. Our study collected clinicopathologic characteristics that may be related to postoperative recurrence, including age, sex, history of hepatitis B, AFP, alanine aminotransferase (ALT), aspartate aminotransferase (AST), γ-glutamyl transpeptidase (GGT), total bilirubin (TBIL), albumin (ALB), platelets (PLT), Child–Pugh class (A or B), BCLC stage, liver cirrhosis, microvascular invasion (positive or negative), and pathological differentiation. Laboratory indicators, such as AFP, ALT, AST, GGT, TBIL, and ALB levels were classified into categorical variables based on a threshold value.

MRI examination was performed in our hospital using 3.0-T MRI systems (Siemens Medical Solutions, Munich, Germany; GE Medical Systems, Chicago, IL, USA; Philips Medical Systems, Best, The Netherlands) within 1 month before surgery. Our liver multiparameter MRI protocol included axial dual-echo (in-phase and opposed-phase) T1-weighted imaging, axial fat-suppressed T2-weighted imaging, diffusion-weighted imaging (DWI, b values of 0 and 800 s/mm²), apparent diffusion coefficient, and liver acceleration volume acquisition (LAVA-XA)-enhanced scanning. Dynamic contrast-enhanced (DCE) LAVA images were acquired at 15–20 s (arterial phase, AP), 50–55 s (portal venous phase), and 85–90 s (delayed phase) after contrast agent injection.

All MR images were independently reviewed by two radiologists with 3 and 10 years of experience to assess the imaging features of HCC. Discordance between the two radiologists were resolved by discussion to consensus. Radiologists were aware that the lesions were HCCs but were blinded to all other clinical data. When multiple lesions were observed, the largest lesions were analyzed. When there were more than two tumors, we defined it as 2; otherwise as 1. The imaging features were selected according to the LI-RADS 2018, including the major features [non-rim arterial phase hyperenhancement (APHE), non-peripheral washout appearance (washout), enhancing capsule appearance (capsule), size, tumor number (num)], and ancillary features (mild‒moderate T2 hyperintensity, restricted diffusion, signal homogeneity, nodule in nodule, mosaic architecture, intratumor hemorrhage, fat in mass, peritumor enhancement during mid-arterial phase, margin, etc.) (10). Our research looked at all nodules’ tumor margins to identify which had smooth margins (defined as nodular tumors in all imaging planes) and which had non-smooth margins (defined as non-nodular tumors in all imaging planes) (11, 12). The maximum cross-sectional diameter of the tumor was measured (avoiding measurement in the AP and DWI due to lesion size overestimation).

Eligible patients were divided into two groups: the training cohort (n = 346) and the testing cohort (n = 148) based on the time of surgery at a ratio of 7:3. We first used the RSF model to select features most relevant to the survival status of the training cohort. RSF uses a collection of decision trees to rank variables based on their importance for time-to-event, which is suitable for reducing the data dimension of highly correlated data. Subsequently, multivariate Cox regression analysis was performed using the stepwise method to identify the independent risk factors. Then, we formulated two nomograms to predict postoperative survival status based on the results of the RSF and multivariate Cox regression analysis, which can provide the clinician with a quantitative tool to predict the individual probability of OS and RFS, respectively. The concordance index (C-index), defined as the proportion of all assessable and orderly patient pairs whose predictions were line with the outcomes, was used to assess the nomogram’s discrimination abilities. To assess the nomogram’s capacity to calibrate, calibration curves were plotted. The nomogram’s clinical usefulness was demonstrated using decision curve analysis (DCA). The testing cohort was used to test and verify the model. A total risk score was calculated for each patient in the entire cohort, based on the points given for each factor in the nomogram. The patients were classified into the high-risk group and low-risk group according to the risk score using recursive partitioning. The Kaplan‒Meier method was used to compare whether the survival distributions differed between the two risk groups. OS curves of different treatments for patients who had recurrence within 2 years after curative surgery were depicted and compared using the Kaplan–Meier method and the log-rank test.

Statistical analysis was performed using SPSS version 22.0 software (Chicago, IL, USA) and R software (version 4.0.4). Categorical variables and continuous variables were compared between the training and testing cohorts using the chi-square test or Fisher’s exact test, and Student’s t -test or the Mann–Whitney U test, respectively. The RSF model analysis and Cox regression analysis were based on the Random Forest SRC and survival package, respectively. Then, stepwise regressions based on the Akaike information criterion were used to further select the above identified variables to construct multivariable Cox regression models. The nomograms were constructed to predict early recurrence and OS. The rms package was used to create the nomogram and calibration curve, while the dca R package was used to perform DCA. The function predictSurvProb() in the pec package was used to extract the survival probability predictions. The C-index was used to assess the nomograms’ prediction performance, compared by and the survival package. All reported P values were two-sided, and P  <  0.05 was considered significant, unless stated otherwise.

For the 5-year OS, multivariate Cox regression analysis based on the RSF model indicated age, AST, BCLC stage, differentiation, maximum tumor diameter, microvascular invasion, and non-smooth margin involvement as independent prognostic variables. The detailed results of the multivariate analysis are shown in Table 2.

For the 2-year RFS, multivariate analysis revealed that BCLC stage, maximum tumor diameter, microvascular invasion, and non-smooth margin were independent prognostic factors. The detailed results of the multivariate analysis are shown in Table 3.

The prognostic nomogram for predicting 5-year OS in patients after hepatectomy is shown in Figure 1A. We constructed a nomogram based on the following seven independent prognostic factors: BCLC stage, non-smooth margin, maximum tumor diameter, microvascular invasion, age, AST, and differentiation. The maximum tumor diameter had the greatest impact on the OS of HCC, according to the nomogram. Users can get a precise likelihood for 1-, 3-, and 5-year survival by adding up the separate scores and calculating the total score. Higher total scores were associated with a worse prognosis. The C-index of the nomogram for predicting OS in the training cohort was 0.787 [95% confidence interval (CI), 0.726‒0.849]. The calibration curves for 1-, 3-, and 5-year OS in the training cohort presented an optimal agreement, as shown in Figures 1B‒D. For the testing cohort, the nomogram of predicting OS also exhibited a high accuracy, with a C-index of 0.711 (95% CI, 0.617‒0.805). The calibration curves also showed favorable calibration of the nomogram for predicting 1-, 3-, and 5-year OS in the testing cohort (Figures 1E‒G). DCA showed that this nomogram added benefit to predicting 5-year OS as compared with treating all patients or treating none of the patients in the training and testing cohorts (Figures 2A‒F).

The prognostic nomogram for 2-year RFS prediction of patients after hepatectomy is shown in Figure 3A. The C-index of the nomogram for 2-year RFS prediction in the training cohort was 0.717 (95% CI, 0.663‒0.770). For the testing cohort, the nomogram of 2-year RFS prediction also exhibited a high accuracy with a C-index of 0.701 (95% CI, 0.616‒0.784). The calibration curves had good agreement for predicting 1- and 2-year RFS in the training and testing cohorts between the predicted and observed survival probabilities (Figures 3B‒E). DCA showed that using this nomogram to predict 2-year RFS was more beneficial than treating all patients or treating the non-patient scheme in the training and testing cohorts (Figures 4A‒D).

We further compared our model with the current BCLC staging system. The results showed that the C-index of the BCLC system was 0.67 (95%CI, 0.67‒0.72) and 0.67 (95%CI, 0.59‒0.75) for OS (Table 4), and 0.64 (95%CI, 0.59‒0.68) and 0.67 (95%CI, 0.60-0.74) for RFS (Table 5) in the training and testing cohort, respectively, which were lower than the C-index of our model (C-index = 0.787, 0.717, respectively for OS and RFS in the training cohort). The DCA for the nomogram and BCLC staging systems are shown in Figures 2 and 4 in the training (A‒C) and testing cohorts (D‒F), respectively. The decision curve showed that using the nomogram to predict OS added more benefits than when using the BCLC staging system when the threshold probability of a patient over 5%, which was comparable in this range.

According to the scoring of nomogram developed with training group, the external cohort 1 was only analyzed for RFS based on follow up data, and the external cohort 2 was validated for RFS and OS. The C-index of the nomogram and BCLC system in the external cohort 1 for RFS was 0.716 (95%CI, 0.605-0.827, p=0.037), 0.627 (95%CI, 0533-0.722), respectively (Table 5). The C-index of the nomogram and BCLC system in the external cohort 2 for RFS was 0.729 (95%CI, 0.613-0.844, p=0.001), 0.587 (95%CI, 0.501-0.673), respectively (Table 5). The C-index of the nomogram and BCLC system in the external cohort 2 for OS was 0.748 (95%CI, 0.623-0.878,p=0.005), 0.624 (95%CI, 0.532-0.717), respectively (Table 4). Besides, the calibration curves had good agreement for predicting 1-, 2-year RFS and 3-,5-year OS in the external cohort between the predicted and observed survival probabilities (Figure 5).

The optimal cutoff point for the prognostic stratification classification was calculated using X-tile 3.6.1 software (https://medicine.yale.edu/lab/rimm/research/software/) in the training cohort, which were 177.7 and 83.7 for OS and RFS, respectively. Accordingly, patients with OS were divided into two subgroups: a low-risk group (total score ≤ 177.7) and a high-risk group (total score > 177.7). The Kaplan‒Meier survival analysis showed that the low-risk group was far removed from the high-risk group for OS in the training cohort, with the low-risk group demonstrating better survival than the high-risk group (Figure 6A). Similar results were observed in the testing cohort (Figure 6B). Correspondingly, patients in the present study were separated into two subgroups for RFS: a low-risk group (total score ≤ 83.7) and a high-risk group (total score > 83.7). The Kaplan‒Meier survival analyses for RFS are shown in Figures 7A, B for the training and testing cohorts, respectively. RFS outcomes were similar to those of OS.

186 patients (37.6%) had tumor recurrence during the 2-year follow-up after hepatectomy in the whole cohort. Of the 186 patients with tumor recurrence, 49 received supportive care, 13 underwent surgery, and 124 underwent transarterial chemoembolization (TACE)/radiotherapy/radiofrequency ablation.

We analyzed retreated OS (reOS), which is the time period from the date of recurrence to the date of death or the last follow-up, to see how different treatments affect recurrent tumors. Kaplan‒Meier survival analysis for reOS is shown in Figure 8 and demonstrated that patients with hepatectomy had a higher reOS than those who received comprehensive treatments and supportive care, while patients with palliative care had the worst prognosis after tumor recurrence.